Libpcap performance under VMWare guest OSes

["Mark Bednarczyk" <voytechs () yahoo com>]

Hello,
    We are running into some performance issues with libpcap when running
under VMWare products such as VMWworkstation or VMServer. These are guest
OSes (running as a VM process on a host operating system.)

We are finding that libpcap drops packets with only few hundred kilo-bits
traffic captured. The same performance tests do not drop any packets even
under many mega-bits/sec throughput for the same host operating system.
"Guest" means under virtual VM machine, while "host" means native OS not
under virtual OS.

Has anyone here worked and figured out if there is anything that can be done
in order to improve libpcap performance under VM products? Here are some
statistics that I've collected on a ubuntu guest OS running in VMWorkstation
under WinXP host.

The test is to run pcap_loop(10000...) 10 times for a total of 100,000
packets. We time and count both packets and bytes transferred and report
these statistics in between pcap_loop calls.

Opening eth2 interface (guest OS on WinXP host):
#000000: recv=10027, drop=17, ifDrop=0 pps=63.90 bps=214.65Kb
#010000: recv=20042, drop=31, ifDrop=0 pps=473.13 bps=3147.04Kb
#020000: recv=30045, drop=31, ifDrop=0 pps=1070.32 bps=7515.12Kb
#030000: recv=40061, drop=49, ifDrop=0 pps=1100.96 bps=7925.73Kb
#040000: recv=50068, drop=57, ifDrop=0 pps=1053.07 bps=7486.51Kb
#050000: recv=60075, drop=63, ifDrop=0 pps=1116.57 bps=7967.32Kb
#060000: recv=70091, drop=78, ifDrop=0 pps=1155.00 bps=8057.36Kb
#070000: recv=80094, drop=80, ifDrop=0 pps=1234.72 bps=8688.54Kb
#080000: recv=90097, drop=85, ifDrop=0 pps=1230.92 bps=8621.55Kb
#090000: recv=100106, drop=94, ifDrop=0 pps=1237.01 bps=8697.74Kb

While the same operating system as host

Opening eth2 interface (Ubuntu host OS no Vmware):
#000000: recv=10000, drop=0, ifDrop=0 pps=1734.00 bps=20333.88Kb
#010000: recv=20002, drop=0, ifDrop=0 pps=1750.09 bps=20518.50Kb
#020000: recv=30004, drop=0, ifDrop=0 pps=1734.91 bps=20340.75Kb
#030000: recv=40004, drop=0, ifDrop=0 pps=1696.93 bps=19892.15Kb
#040000: recv=50002, drop=0, ifDrop=0 pps=1732.20 bps=20302.28Kb
#050000: recv=60004, drop=0, ifDrop=0 pps=1702.13 bps=19953.75Kb
#060000: recv=70003, drop=0, ifDrop=0 pps=1729.80 bps=20275.08Kb
#070000: recv=80001, drop=0, ifDrop=0 pps=929.02 bps=10889.94Kb
#080000: recv=90001, drop=0, ifDrop=0 pps=1715.56 bps=20111.15Kb
#090000: recv=100004, drop=0, ifDrop=0 pps=1737.92 bps=20373.84Kb

The disproportionate bit rates are due to the fact that 1st test was a
torrent download over a slower internet connection, while the second was a
host to host copy on local LAN. None the less you can see that under
VMWorkstation, libpcap begins to drop packets after just a few Kilo-bits/sec
data rate, while under native host OS even at 20Mbps, the there were no
drops. It's a lot worse when a pcap filter is applied.

Tests were done under VMWorkstation version is 6.5 and, VMServer 4. Libpcap
0.8 (ubuntu, but also under deb5, FC11 and others).

Project: http://jnetpcap.com - a libpcap java wrapper. I have taken java out
of the picture for these tests, by using empty native callback handler for
pcap_loop and pcap_dispatch calls.

Cheers,
mark....

"It is impossible to grasp the meaning of the idea of sound money if one
does not realize that it was devised as an instrument for the protection of
civil liberties against despotic inroads on the part of governments.
Ideologically it belongs in the same class with political constitutions and
bills of rights."  Ludwig von Mises


-
This is the tcpdump-workers list.
Visit https://cod.sandelman.ca/ to unsubscribe.